An understanding reply is only the beginning

AI can explain psychological ideas, help you find words and offer a response when you feel alone. But an answer that sounds understanding does not establish that you are receiving safe, suitable therapy. A chatbot’s warmth, confidence and availability can make that distinction surprisingly easy to lose.

Therapy involves a relationship in which your experience can be explored, misunderstood, checked and understood more fully over time. What matters includes how another person responds to you at a particular moment, and what happens between you when the response does not fit.

If you have been using AI because there are things you find difficult to tell another person, that can be a useful place to begin counselling with me. We can explore what those conversations have offered, what they have missed and what you need from a human relationship. You do not have to defend having looked for support.

Why people turn to AI for therapy

Imagine being awake at two in the morning, turning the same worry over again. Your friends are asleep. You may be waiting for support, unable to afford regular sessions, or unsure whether your problem is serious enough to bring to someone. A chatbot offers an immediate place to put it.

There is no unfamiliar person watching while you search for the words. You can rewrite a sentence, ask an awkward question or close the conversation. For someone who fears being judged, that sense of control may make disclosure feel possible.

A 2025 study of people already using ChatGPT for emotional or mental health support found uses ranging from expressing distress and navigating relationships to learning about mental health. Participants described benefits as well as superficial emotional engagement and limitations in the information offered. These were accounts from selected users, not evidence that everyone benefits or that a chatbot provides effective treatment.

The appeal deserves to be taken seriously. So do the gaps in access that make a readily available conversation so attractive. Telling someone simply to stop using AI does not create a therapist they can afford, an appointment they can attend or a person they feel safe enough to trust.

The words can sound caring without a person caring

“That sounds really painful” can be a helpful thing to read. It can also be an easy sentence for a system to produce. Its emotional effect on you may be real, while the process producing it is very different from another person trying to understand your life.

A chatbot can recognise patterns in what you write and generate language that resembles empathy. That does not give us a basis for assuming it feels concern, shares a human experience of vulnerability or understands you in the way a person can. Feeling comforted by the words and attributing human feelings to their source are different things.

When a chatbot says, “That must feel painful,” you may be left with the pain, but without the human support you were reaching for. Having the feeling named can bring relief, yet it can also leave you feeling more alone, as though you have been abandoned in the very pain you have just disclosed. A response has arrived, but there is no person there to be moved by what you are sharing and stay alongside you.

A human therapist can convey empathy through their eyes, body language and tone of voice, as well as through words. The gentleness of a look, a change in their voice or a quiet willingness to remain with you can help communicate, “I am here with you.” AI can generate reassuring language and simulate a caring voice, but it cannot offer that shared, embodied human presence. Therapy does not always take pain away. Part of what it can offer is support while you feel it.

In this hypothetical example, you tell a therapist, “I’m fine with the separation,” then fall silent. Perhaps your voice has changed. Perhaps you smile while your eyes fill. A human therapist can be affected by that moment and gently ask whether “fine” is the whole story. They can also ask whether you would rather sit quietly than explain.

A text chatbot receives the words you enter. Systems with voice or video may receive more cues, but detecting a pause or an expression does not itself establish human understanding. More information about someone is not the same as sharing an encounter with them.

Human therapists get this wrong too. I cannot know exactly what you need from a sentence or a silence. I can notice, ask, listen to your correction and take responsibility for how I respond. In person-centred work, your account of your experience matters more than my confidence about an interpretation.

You can change the tone. That does not make the answer right

You can ask many chatbots to be gentle, reassuring, blunt or more challenging. That flexibility can make an explanation easier to take in. It can also create the impression that you have found the right kind of therapist when you have mainly changed the style of the reply.

Consider another hypothetical example. You describe a difficult evening with your partner and ask, “Am I right that they don’t care about me?” A sympathetic answer may leave you feeling supported while giving too much weight to your first interpretation. Asking the system to “be brutally honest” could produce a harsher judgement without making it better informed.

Research published for CHI 2026 found that adding information from previous user interactions often increased excessive agreement in the models tested, although the effect varied. This was research on model behaviour, not a therapy trial. It does raise a useful question: does a reply help you examine an assumption, or mainly make the assumption feel more certain?

A human therapist should not simply agree with you either. Nor should therapy mean paying someone to criticise you. We can take your hurt seriously while remaining curious about what happened, what is uncertain and what your response might mean for you. The challenge needs to fit the person and the moment, rather than a tone setting.

Could we ever form a human bond with AI?

I think the question of replacement reaches beyond how convincingly a system talks. It concerns whether AI comes to be accepted into ordinary human life as a relational presence, and whether people experience a bond with it that feels human.

Some people already describe trust, closeness and a sense of being understood by AI. We should not dismiss those feelings because the relationship is with software. But a bond felt by one person does not, on its own, show mutual emotional experience, clinical competence or a safe basis for therapy.

A 2025 randomised trial of Therabot, a purpose-built mental health chatbot, reported improvements in symptoms compared with a waiting group. Users also reported a strong therapeutic alliance, a sense of trust and collaboration. The research team evaluated conversations and was prepared to intervene over safety concerns. This was a particular system under studied conditions, not an ordinary chatbot proved equivalent to a human therapist. The trial did not directly compare the two. The research team’s account is useful context.

That is promising evidence, and it complicates any easy claim that people cannot connect with AI. It does not settle what that connection provides, what it lacks or what happens over a much longer period. Social acceptance may change. Whether it will bring a relationship equivalent to human therapy remains an open question.

For me, being attached to a responsive system is not sufficient grounds for treating it as a replacement for a person who can meet you, be affected by you and be accountable for the work.

What else can get lost in the conversation?

A fluent explanation can be persuasive before you have had a chance to question it. If a chatbot offers a psychological label for you, your partner or your childhood, that label may begin to organise how you see the situation. A convincing account still needs evidence and context. It is not an assessment simply because it sounds professional.

There are safety concerns too. In a 2025 experimental study of language models and therapy chatbots, researchers found stigmatising responses and inappropriate replies to some high-risk scenarios. The findings concern the systems and situations tested, rather than every product now available. They nevertheless show why an apparently caring conversation cannot be taken as proof that a system will respond appropriately when the stakes change.

Privacy is another part of the relationship. Before sharing intimate material, look at who provides the service, how conversations are stored and used, and what control you have over them. A friendly interface does not tell you those things. Do not assume a chatbot conversation has the same confidentiality arrangements as counselling, and remember that other people’s private messages contain their information too.

There is also a personal question worth noticing: does using the chatbot help you return to your life, or is it becoming the place you repeatedly go to postpone a difficult conversation? A tool can help you find the words. You may still need a relationship in which to say them.

Therapy has boundaries, confidentiality arrangements and professional responsibilities that we can discuss openly. Those do not make a human therapist infallible. They give you something more substantial to ask about than whether the latest answer felt reassuring.

Where AI can be useful

Psychoeducation means learning about psychological experiences and approaches. AI can be a useful starting point for this: asking for a plain-language explanation of avoidance, exploring the difference between a thought and a fact, or finding questions you would like to bring to a session.

It may help you put an experience into words or organise a reflection when a blank page feels difficult. You can ask for a shorter explanation or an everyday example. The value lies in what you can then understand and consider, rather than in accepting the response as a verdict about you.

Check important information against the original research or a reliable health source. If the system supplies a reference, check that it exists and supports the claim. An explanation of a condition cannot establish that you have it, and a suggested exercise is not automatically suitable for your history or circumstances.

When you are using these ideas to work on distress or persistent difficulties, applying them alongside a therapist gives room for the context an explanation can miss. Something presented as a helpful challenge might feel overwhelming. Advice to set a boundary might overlook a relationship in which doing so is unsafe. Together we can consider what fits, what needs adapting and what to leave aside.

You do not need a therapy appointment to ask an educational question. But when information becomes a plan for treating yourself, or instructions for a major decision, the distinction matters.

One useful question to bring into a session

If you have used AI for support, you might ask yourself: “What did I hope this conversation would give me?” Perhaps it was reassurance, permission to feel angry, a clear explanation or the experience of someone staying with a difficult feeling.

You can bring that hope into counselling, along with one idea that helped or unsettled you. You do not need to hand over an entire private chat. We can explore what the response meant to you and what you were looking for underneath it.

The aim is to strengthen your own understanding and choices, including noticing the qualities and relationships you can draw on. You do not have to outsource your judgement to a chatbot, or to a therapist.

Questions people ask about AI therapy

If AI helps me feel better, does that count?

Yes, the relief you feel matters. You can value it while still asking whether the information is sound, whether the conversation helps over time and what support you may need beyond it. Feeling better after a reply does not answer all of those questions.

Can I use AI between therapy sessions?

You can talk with your therapist about how you use it, what you hope it offers and what happens afterwards. General learning or organising questions may be useful; advice that contradicts your work or leaves you more distressed deserves attention. Agreeing how to approach this can make its role clearer.

Will you judge me for talking to a chatbot?

No. I would want to understand what brought you there and what the experience has been like for you. Looking for support, particularly when access is difficult, is something we can meet with curiosity and care.

A conversation with someone who can meet you

If you have found yourself telling AI things you have not felt able to say to anyone else, you are welcome to begin there with me. You do not need to arrive with a diagnosis, an explanation or a decision about whether the chatbot was helpful.

In individual counselling, we can explore the loneliness, uncertainty, shame or relationship difficulties behind those conversations, as well as what is already helping you cope. My person-centred and ACT-informed approach offers room to understand your experience, make space for difficult feelings and consider choices that connect with what matters to you.

I work in Shoeburyness, Southend-on-Sea, and online across the UK. You can book a free introductory conversation to talk about what you are looking for and whether working together feels right. There is no obligation to continue.

You may already have an explanation for what you are going through. You can still want someone to be there while you go through it.

Sources and related reading

Luo, X., and colleagues (2025). Seeking Emotional and Mental Health Support From Generative AI: Mixed-Methods Study of ChatGPT User Experiences. JMIR Mental Health. User accounts, rather than a clinical trial.

Jain, S., Park, C., Viana, M., Wilson, A., and Calacci, D. (2026). Interaction Context Often Increases Sycophancy in LLMs. CHI 2026. Open author manuscript.

Heinz, M. V., Mackin, D. M., Trudeau, B. M., and colleagues (2025). Randomized Trial of a Generative AI Chatbot for Mental Health Treatment. NEJM AI. A trial of Therabot with a waiting-group comparison, rather than a direct comparison with human therapists.

Moore, J., and colleagues (2025). Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers. ACM FAccT 2025. Experimental evaluations of the systems tested.

For a related reflection on what screens can miss between people, read Texting Your Partner: How Messages Can Drive You Apart. You can also explore how I work in counselling.